NVIDIA robots achieve 99% success via self-improving AI loop

What happened

NVIDIA has developed ENPIRE, software that lets physical robots autonomously improve their own performance through a feedback loop similar to AI agents. The system includes automatic evaluation, reset, policy refinement, and code improvement modules. Coding agents using the framework achieved a 99% success rate on dexterous manipulation tasks such as organizing pins and cutting zip ties in the real world.

Why it matters

The system minimizes human effort by automating evaluation and reset—tasks that historically required constant human supervision. This suggests a path toward robots that can self-improve without human intervention, which could reshape manufacturing and assembly work. However, the approach currently works best on simple, well-defined tasks where automatic evaluation and reset are feasible.

What to watch

Performance varies by AI model—GPT-5.5 and Opus 4.7 trade advantages, while larger agent teams (e.g., 8 agents) find better solutions faster than single agents. Each workstation runs an NVIDIA RTX 5090. The main scaling challenge is that robot resources are underutilized when agents read logs or debug, so adding more robots does not naturally parallelize.